plat 203

← Pulse platform issue index

PLAT-203 — pi-cli-routed models had no rate card at all, and a genuinely-zero cost was indistinguishable from "we don't know"

Coordination	Value
Assigned agent	unassigned
Ticket state	fixed
Last synchronized	2026-08-29

• Priority: P2 — cost figures silently absent (not zero) for every pi-cli-routed call, which "could mislead operations review into under-counting cycle cost for any step pinned to a non-Claude model" per the finding's own impact assessment.
• Owner: agent_go/pkg/orchestrator/base_orchestrator_tokens.go (calculatePricingFromModelData), agent_go/pkg/orchestrator/cost_storage.go (buildModelTokenUsage, EnsureModelTokenUsagePricing), agent_go/pkg/orchestrator/base_orchestrator_types.go (ModelTokenUsage).
• Related: harness:cost-ledger:reflection-attribution-and-call-count (confida-login, medium; classification: "no_issue", recommended_route: "evidence_wait" in the harness's own self-assessment — this ticket found it was in fact a real, confirmed gap). Same class of zero-vs-unset ambiguity costledger.Aggregate (the newer, parallel cost-tracking system — see PLAT-184) already solved via BillingBasis="unpriced"/UnpricedCallCount; this ticket applies the same pattern to the older, still-live ModelTokenUsage/cost_storage.go system the finding's cited file (costs/execution/<workflow>/<date>.json) actually uses.

The finding

costs/execution/confida-staging/2026-08-24.json's by_step_and_model for execute-browser-and-capture-apis showed google/gemini-3.7-flash with llm_call_count=4 present but no total_cost_usd key at all in that model block.

Root cause — two compounding gaps, both confirmed in code

1. pi-cli has no rate card at all, for any model. picli.PiCLIAdapter.GetModelMetadata (multi-llm-provider-go/pkg/adapters/picli/picli_adapter.go:90-104) returns *llmtypes.ModelMetadata successfully (no error) but never sets any *CostPer1MTokens field — not specific to gemini-3.7-flash, every model this adapter serves gets the same empty pricing. Confirmed gemini-3.7-flash is in fact the adapter's own current default model (DefaultModelID, picli_adapter.go:13), so this isn't a stale/missing entry for an unusual model — it's a structural gap for the whole provider.
2. A genuinely-zero cost and an unknown cost were indistinguishable. ModelTokenUsage.TotalCost had json:"total_cost_usd,omitempty" — when calculatePricingFromModelData fell through to all-zero costs (metadata missing, or present but empty as in case 1), the JSON key simply vanished, identical to what a genuinely free call would produce.

Fix

calculatePricingFromModelData now returns a pricingFound bool: true only when the resolved model metadata actually carries a nonzero rate on at least one of Input/Output/Reasoning/CachedInput/CachedInputWrite cost-per-1M-token fields — not merely "metadata lookup didn't error."

ModelTokenUsage gained Unpriced bool (omitempty — a priced call's JSON is byte-identical to before this change) and TotalCost lost its omitempty (a priced call already always has a real total; an unpriced call now explicitly shows "total_cost_usd":0,"unpriced":true" instead of a missing key). Both call sites (buildModelTokenUsage, EnsureModelTokenUsagePricing) set Unpriced from pricingFound and log a [COST] line naming the exact provider/model with no rate card, so a recurrence is immediately debuggable instead of silent.

Explicitly not done

• Did not add real per-token pricing for google/gemini-3.7-flash or any other pi-cli model. That requires knowing pi-cli's actual billing model (subscription passthrough vs. metered vs. genuinely free tier) with confidence — guessing a dollar figure would be worse than an honest "unpriced" marker, since a plausible-looking wrong number is harder to catch than an explicit unknown. This is now a well-scoped follow-up if pi-cli's billing terms are confirmed.
• Did not change the update-condition in EnsureModelTokenUsagePricing (still gates writing pricing fields on inputCost>0 || ... || totalCost>0, not on pricingFound) — changing that would also affect a real rate-card model whose token counts happen to be zero, which no existing test covers; scoped this change to exactly what the finding needed.
• Did not touch costledger/costobserver (the newer, parallel system) — it already has its own equivalent BillingBasis/UnpricedCallCount disambiguation and was not what the finding's cited file used.

Verification

• go build ./... clean.
• New tests TestUnpricedProviderCallsAreExplicitNotAbsent and TestPricedProviderCallsAreNotMarkedUnpriced (pkg/orchestrator/unpriced_model_cost_test.go) prove: a pi-cli/gemini call is marked pricingFound=false, Unpriced=true, and its JSON explicitly shows "total_cost_usd":0,"unpriced":true; a real claude-sonnet-5 call keeps pricingFound=true, Unpriced=false, a real nonzero total, and — critically — no unpriced key at all in its JSON, so every existing priced-call consumer sees a byte-identical payload to before this change.
• Existing TestPhasePricingUsesImmutableClaudeModelRateCards (Claude rate-card exactness) passes unchanged.
• Full suite: 3 pre-existing failures before and after (confirmed via git stash), all in cmd/server/guidance, none touching this ticket's packages — no regression.